Hi, long time user first time posting 👋 I have a question about the meta-omics pipelines. One particular approach that I can't seem to find in any of the pipelines is how to obtain the functional profile of a metagenome/metatranscriptome without assembling it first. In other words, something that maps unassembled reads to a database of proteins/genes. The first tool that comes to mind is HUMAnN, though other exists. Is there any plan to include this on any of the pipelines? Is there a reason why the developers have favored an assembly-based approach?

Agreed. I made myself https://github.com/d4straub/functionalmetagenomics because I had no time for a proper nf-core pipeline 😄

There is #C01U1E95JQ5 coming up. The idea is to map to genomes, but there's no particular reason for why you can't use it to map to a set of genes except that currently were mapping to contigs and extracting counts for each gene with a gff file.

Same with taxprofiler, wanted the scoep was to stay: 'who is there' I think the thing with HUMAnN too is there are not many 'competitors' to be worth pieplining... But loads of people have requested it, so I think a standalone pipeline might be worth it if anyone is interested 🙂

https://github.com/KoslickiLab/fmh-funprofiler There’s also this one which uses a kmer sketching approach - haven’t tried it myself though. I also built my own functional profiling pipeline (https://github.com/prototaxites/mgannotate) but this also uses an assembly based approach. In my case, that’s because I already had the assemblies to hand, so it seemed silly to not use them for gene prediction. Code is a little bit scratchy as I was modifying it as I went during analysis - so not necessarily recommended for use. I guess the thing about a non-assembly approach is that if you have enough data, I imagine an assembly approach is more accurate and will give you de novo genes, rather than just mapping to known sequences. So in general that is probably preferable if you have the resources?

if you have enough data, I imagine an assembly approach is more accurate

I also think so. But "enough data" is not always there, when economic constrains are playing a role.

This also depends on whether one's samples mostly contain sequences very similar to what's found in databases or not. In #C01U1E95JQ5 -- at least in one mode -- we're banking on microbial communities becoming quite well covered by genome (MAG) projects, to allow mapping. If one can tap into functional annotations and curated databases that provide a depth, likely for specific questions, that is not provided by general tools, I think mapping reads can be very successful even for large datasets. I'd be very interested in implementing such databases and tools in e.g. #C01U1E95JQ5 and #C02FC6VFQG1, but a stand alone pipeline is probably also good.

Thanks all for the discussion. I do agree that assembly-based methods can be very powerful but as said, mapping has its place as well. I'm thinkin in particular on the analysis of human microbiomes (which will likely be the most common case) at shallow sequencing depths, which are pretty cost-effective to get a broad overview of taxonomic and functional composition, with lower computational requirements. I also keep thinking of the reproducibility or the comparability of a functional profile obtained by mapping to a coassembly. If after running a first batch of samples I get more, I'd need to re-run all samples to make them directly comparable, right? This wouldn't be the case with a mapping based approach (or if samples were assembled independently). In any case, I am very interested in having something like this available. Not sure I have the skills to implement, but I'm more than happy to test and help document if needed.

I think there might be two different things here, that are both interesting to us as a community. First is a pipeline that focuses on functional annotation of reads. But that will depend on various annotation sources -- databases and tools -- that can be valuable in other pipelines. Nice things to think about for the hackathon maybe?

@James Fellows Yates regarding tools other than HUMAnN, I've seen https://github.com/KoslickiLab/fmh-funprofiler and https://bitbucket.org/bromberglab/mifaser/src/master/. Surely the folks here are aware of others available

I think you have a good candidate for a pipeline there @Jacobo de la Cuesta!

Is there a reason this kind of functional profiling isn’t implemented in #C02K5GX2W93, alongside the more specific AMR gene finding etc.?

scope: funcscan is for contigsor genomes not (short) reads